Full Report
Cybercriminals are hijacking artificial intelligence (AI) user accounts via information stealer logs to create "stolen keys" that grant illicit access to tools from model providers like Google, Anthropic, and others. Information stealers like Lumma Stealer or Vidar are equipped to harvest a wide range of data from compromised systems. This can include credential, session tokens, and API
Analysis Summary
# Tool/Technique: Infostealer-Enabled AI Session Hijacking (LLMjacking)
## Overview
This technique involves the use of information-stealing malware to harvest authentication secrets—specifically JSON Web Tokens (JWTs), JSON Web Encryption (JWE) structures, and API keys—from compromised systems. Threat actors replay these stolen secrets to bypass Multi-Factor Authentication (MFA) and gain illicit access to premium Artificial Intelligence (AI) and Large Language Model (LLM) services like Google Gemini, OpenAI, and Anthropic.
## Technical Details
- **Type:** Malware Family (Infostealers) & Technique (Session/Token Replay)
- **Platform:** Windows, macOS (targets browser data and local storage)
- **Capabilities:** Credential harvesting, session token extraction, API key discovery, MFA bypass via replay.
- **First Seen:** Reported significantly in September 2026 (based on August 2026 data dumps).
## MITRE ATT&CK Mapping
- **TA0006 - Credential Access**
- T1539 - Steal Web Session Cookie
- T1552.001 - Unsecured Credentials: Private Keys
- **TA0001 - Initial Access**
- T1078 - Valid Accounts
- **TA0004 - Privilege Escalation**
- T1550.004 - Use Alternate Authentication Material: Web Session Cookie
- **TA0010 - Exfiltration**
- T1041 - Exfiltration Over C2 Channel
## Functionality
### Core Capabilities
- **Browser Data Harvesting:** Extraction of cookies, saved passwords, and autofill data from popular browsers.
- **Token Extraction:** Specifically targeting JSON Web Tokens (JWT) and NextAuth.js JWE structures that manage sessions for AI providers.
- **API Key Scraping:** Scanning local files and environment variables for plaintext API keys (e.g., OpenAI, Groq, OpenRouter).
### Advanced Features
- **Session Replay:** The ability to use stolen JWTs to impersonate a user without needing a password or passing an MFA challenge, as the token represents an already authenticated state.
- **LLMjacking:** Weaponizing stolen API keys to utilize high-compute AI resources at the victim's expense, often for automated spam, malware generation, or resale.
## Indicators of Compromise
- **File Hashes:**
*(Specific hashes for the 7GB dump were not provided in the article, but typical for these families:)*
- **Lumma Stealer:** Varies due to frequent packing.
- **Vidar:** Varies by campaign.
- **File Names:** `logs.zip`, `gram_data.rar` (Common in Telegram-based stealer log trades).
- **Network Indicators:**
- C2 communication typically involves encrypted traffic to domain-generated algorithm (DGA) addresses or Telegram API endpoints.
- **Behavioral Indicators:**
- Unauthorized access to LLM accounts from anomalous IP addresses.
- Unexpected spikes in API usage/billing for AI services.
- Browsers attempting to access `Local State` or `Login Data` files via non-browser processes.
## Associated Threat Actors
- Users of **Lumma Stealer**
- Users of **Vidar**
- General "Log Divers" (Cybercriminals who purchase and sift through bulk stealer logs for high-value targets).
## Detection Methods
- **Signature-based detection:** Antivirus/EDR signatures for Lumma and Vidar binaries.
- **Behavioral detection:**
- Monitoring for "Impossible Travel" (logins from geographically distant locations in short timeframes).
- Monitoring for sensitive file access (browser profile folders) by unsigned or suspicious processes.
- **YARA rules:** Rules targeting the specific string structures of JWTs or API keys within suspicious memory dumps.
## Mitigation Strategies
- **Device Bound Session Credentials (DBSC):** Implementing Chrome’s DBSC to cryptographically link session tokens to specific hardware, preventing replay on other devices.
- **IP Allowlisting:** Restricting access to AI service API keys and enterprise accounts to known corporate IP ranges.
- **Token Lifespan Reduction:** Shortening the expiration time of JWTs to minimize the window of opportunity for replay.
- **Secrets Management:** Using tools like TruffleHog to scan for and revoke leaked API keys immediately.
## Related Tools/Techniques
- **LLMjacking:** The specific act of stealing AI resources.
- **Cookie Stealing (Pass-the-Cookie):** The broader category of this attack.
- **TruffleHog:** A tool used by both defenders and attackers to find secrets in data dumps.